ComfyUI-FL-RelayVSR

FL RelayVSR

4x real-world video super-resolution for ComfyUI, powered by RelayVSR. A one-step Wan model generates sparse reference keyframes and a compact Flash decoder upscales every frame against them.

Features

• 4x video upscaling - Any IMAGE frame batch in, exactly 4x the resolution out; single images work too
• Automatic model download - The loader fetches the pinned checkpoints from Hugging Face on first run
• ComfyUI memory management - The model loads and offloads through ComfyUI's model manager, freeing other models when VRAM is needed
• Arbitrary input sizes - Frames are padded to the model grid internally and cropped back
• Progress and cancel - Per-frame progress bar; Cancel stops between frames

Nodes

| Node | Description | |------|-------------| | FL RelayVSR Model Loader | Downloads (if needed) and loads the Wan reference model and Flash decoder | | FL RelayVSR Upscale 4x | Upscales an ordered video frame batch 4x |

Installation

cd ComfyUI/custom_nodes
git clone https://github.com/filliptm/ComfyUI-FL-RelayVSR.git
cd ComfyUI-FL-RelayVSR
pip install -r requirements.txt

The Wan and Flash attention kernels are written in Triton. Linux PyTorch ships with Triton; on Windows install the triton-windows build that matches your PyTorch version.

Models

On first execution, FL RelayVSR Model Loader downloads these files from kopper/RelayVSR into ComfyUI/models/RelayVSR:

| File | Size | |------|------| | wan.safetensors | 10.1 GB | | flash.safetensors | 128 MB |

Disable download_if_missing to require files you placed there yourself. No network request happens when both files are present.

Quick Start

1. Load the example workflow examples/relayvsr_video_upscale.json.
2. Pick a video in Load Video and set the Trim Video duration (0 = whole clip).
3. Queue. Frames go through Get Video Components -> FL RelayVSR Upscale 4x -> Create Video, which keeps the source FPS and audio.

Settings

• window_size (default 16): a Wan reference keyframe every window_size - 1 frames. Larger is faster; smaller tracks the references more closely.
• seed: noise seed for the one-step Wan keyframes. Fixed by default so reruns are cached.
• precision: bf16 by default; use fp16 on GPUs without bf16.
• prompt (optional, experimental): replaces the built-in Wan prompt with CLIP Text Encode conditioning from the Wan umT5-XXL text encoder (CLIPLoader type wan). The effect is subtle: in testing it changed only fine edges and texture (about 2/255 mean difference).

Performance

Measured on an RTX PRO 6000 Blackwell (bf16):

• 624x360 to 2496x1440: about 13 FPS once warm, 15.5 GB peak VRAM (9.4 GB of weights plus about 1.8 KB per output pixel).
• 960x540 to 3840x2160: 273 frames in about 79 s end to end, including video decode and encode.
• The first run in a fresh Python environment also compiles the Triton kernels.

Upscaled frames are held in system RAM as a float32 IMAGE batch, which is about 43 MB per 2496x1440 frame. Trim or split long clips.

Requirements

• NVIDIA GPU with CUDA and Triton
• 16 GB+ VRAM for ~1080p output
• Python 3.11+

Credits

• RelayVSR by Xijun Wang, Xin Li, Zirui Lang, Suhang Yao, Haoran Li and Zhibo Chen (paper)
• The inference code in vendor/refvsr_inference is vendored from upstream; see vendor/README.md

License

Apache-2.0, matching the upstream RelayVSR code. Model weights are distributed by their authors on Hugging Face under their own terms.